New 'Harmony' Metric Reveals How Average Scores Hide Uneven Model Performance

DanielKhashabi · x · 2026-08-21

Daniel Khashabi's team proposes "Harmony," a new metric to measure how uniformly a model's competence is distributed across different subdomains of a benchmark. The paper argues that relying on average scores is misleading, as overall accuracy can be disproportionately influenced by specific subdomains (e.g., Biological Concepts in ARC-Easy). They recommend reporting Harmony alongside accuracy to provide a more complete, distributionally reliable view of model performance.

Original post →

More from Research

Research channel →